PTXNet: An extended UNet model based segmentation of pneumothorax from chest radiography images

Abstract

AbstractIn the respiratory analysis of human beings, lungs play an important role and can be used to find several associated diseases. One of such conditions is called pneumothorax which occurs due to leaking of air into pleural space and results in lungs collapse. The traditional methodology uses chest radiography with the manual intervention of radiologists for the detection and segmentation of the affected areas. This paper introduces a deep neural network based methodology for the automatic localization and segmentation of the proper region‐of‐interest (RoI). The proposed approach is based on transfer learning where the existing UNet model is extended and redesigned to a new architecture named PTXNet for RoI segmentation. In PTXNet, the traditional encoder is redesigned with the use of EfficientNet, SE‐ResNeXt50 and Xception convolutional neural network (CNN) architectures. Furthermore, residual blocks are introduced in the decoder phase and concatenation of the previous decoder stage feature maps in addition to standard global skip connections is performed. The PTXNet is trained on a dataset of more than 15,000 chest radiography and resulted in the mean dice coefficient of 84.89%. It is found that the proposed approach provides superior results than the UNet model with an increase in the mean dice coefficient of 18.76%.